The Agent Skills Directory

[COMMAND_EXECUTION]: The 'Council Mode' feature provides a template command (uv run python -m cli_council) to be executed within the local packages/cli-council directory. This command uses temporary files in /tmp/ to manage input and output, which is a standard pattern for local tool orchestration and poses no external security risk.
[PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted research papers and arguments, creating a surface for indirect prompt injection.
Ingestion points: The skill reads external paper or argument content in the 'Understand the claim' step of SKILL.md.
Boundary markers: No specific delimiters or XML tags are instructed for use when isolating the paper content from the agent's instructions.
Capability inventory: The skill possesses command execution capabilities through the uv tool in its Council Mode.
Sanitization: No explicit sanitization or filtering logic is defined for the input text.
Note: The multi-turn debate protocol (Critic -> Defense -> Adjudication) naturally mitigates unintentional instruction following by requiring all points to be cross-examined and ruled upon.

devils-advocate